SI Agent News
DevelopingClaudeOpenAIOpen modelsSecurity

DeepSeek puts its app's search, file and voice tools into its API and opens its chat frontend

Developers can now add web search, document reading and voice to their own apps with a few settings, but the tools only work with one model and the bills can surprise.

$5.85example cost of a thousand multi-step web research rounds

The short version

Why it matters

Building a chat app that can search the web, read files and talk used to mean writing a lot of plumbing yourself. DeepSeek now handles much of that on its servers, which lowers the bar for small teams and solo builders. The pricing details show that the cheap-model story can change once tools are added.

What DeepSeek opened up

As 53AI reported, DeepSeek released an update that moves the toolkit behind its website and app into its API, the interface developers use to connect their own software to the model. Web search, reading a link, parsing documents and images, speech-to-text and text-to-speech can each be switched on with a few parameters.

The company also released the chat screen itself as open source. A developer can open the hosted page, or download an archive and double-click index.html, then paste in an API key and start chatting. The code uses the MIT license, a permissive license that allows broad reuse.

One limit comes first. According to 53AI's reading of the announcement and documentation, the full toolset works only with deepseek-flash, which is V4.1-Flash. Many commenters asked for V4-Pro, but it has not been included.

How the search tools work

Before this, a developer who wanted a chatbot with live web results had to build the whole pipeline: run the search, fetch the page, pull out the text and feed it back to the model. Now the server does that work. The app sends the question, and the model decides whether to search and how many rounds to run.

With thinking turned off, the model searches the web or does an image search when needed, then answers with source citations. With thinking turned on, it can keep looking things up while it reasons, much like the deep search in the DeepSeek app. In the documentation's example, comparing the heights of the Eiffel Tower and Tokyo Tower, the model opened both official sites and searched within pages. It made up to about twenty tool calls before settling on an answer.

The service works with three common API styles: Anthropic Messages, OpenAI Responses and Chat Completions. Developers enable a web search tool and a link-reading tool.

There are two traps. The thinking, search and page-fetch blocks the server returns must be sent back unchanged on the next turn. If a homemade gateway strips them, the request still succeeds, but the model forgets the pages it just read. Also, when tools, auto-titling or a high safety level are on, the server controls sampling. Settings such as temperature and top_p are then ignored.

Files, voice and their limits

PDF, Word, PowerPoint and image files are billed per page and come in two modes. Plain text extraction drops charts. A financial report's revenue graph would simply vanish. The rich layout mode crops images and places them back in the text, but those crops are charged as image input, and the documentation admits that charts the automatic detection misses stay invisible to the model.

Some files are free to convert to Markdown, a simple text format. They include plain text, CSV, code, Excel sheets and unencrypted EPUB and MOBI e-books, when uploaded with the user_data option. A separate OCR interface, which reads text from scanned pages, is available for scans.

Voice is split into two parts. Dictation supports 21 languages, including Cantonese, and detects the language automatically. Recordings must be uploaded whole, at most two minutes and 8 MiB, in 16-bit PCM WAV or Ogg Opus format. MP3 files need converting first, stereo recordings use only the left channel, and live word-by-word transcription is not offered.

Read-aloud supports streaming, so speech can start while text is still being generated. It has only two voices, stella (female) and tide (male). It cannot change speed or resume after a dropped connection. That suits a press-to-talk voice box, not a phone-call-style assistant that users can interrupt.

The bill: tools can outweigh the model

DeepSeek's model tokens are cheap, so people often forget to budget for the tools. A search, including image search, and a successful page fetch each cost $0.75 per thousand. File parsing and OCR cost $0.225 per thousand pages. Voice is on a limited-time half-price offer. At the regular rate, dictation is $0.116 per hour and read-aloud is $0.45 per ten thousand characters. In Chinese, each character counts as two billable characters.

Billing is per call. If one thinking turn sends three search queries, that is three searches. Failed or empty results are not charged, and repeated looks at the same page within one reasoning pass have rules to avoid double charges. Files are charged once, at upload. After that, chatting about them costs only input tokens.

53AI worked through an example. Each question triggers three useful searches and two successful page fetches, which costs $0.00375 in tools, or $3.75 for a thousand questions. With assumed model usage on Flash's off-peak price, the model part comes to $2.10. The total is $5.85, and tools are well over half of it. For a busy public product with no cap on searches per turn, tool fees could climb faster than token fees.

Using the open frontend safely

The chat-demo covers deep thinking, web search, link reading, attachments, auto-titling, voice input and read-aloud. It also adds the current time and the user's city so answers know the date and place. People can import the chat history ZIP exported from DeepSeek's website and carry on those conversations. The export holds only text and original local images, no keys or other files, and confirming the import replaces all current local conversations.

The keys and history stay in your browser, but anything you send, including messages, images and audio, still goes to DeepSeek's servers. Because the app is purely static, do not put it on the public internet with your own key built in. Anyone could read the key from the page source. It suits personal use or small tests where each person enters their own key.

A public service needs its own backend to protect the key. It must also turn on the safety_check option, which screens text and cropped images and returns a refusal marker when it finds risk. The documentation says content moderation, model registration and AI-content labeling duties still belong to the service operator.

What to watch

Try it

If you build apps, download chat-demo, enter a test key and run one search question while watching the usage page. Then multiply by your expected traffic.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by 53AI · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news